CMU-Q 15-381 Lecture 1: Introduction AI, basic definitions, - - PowerPoint PPT Presentation

CMU-Q 15-381

Lecture 1: Introduction – AI, basic definitions, problems, road map

Teacher: Gianni A. Di Caro

OUTLINE

2

§ AI? Representation and Problem Solving? § What this course is about § Fundamental definitions and notions § Road map

AI: IN FICTION

3

In fiction, popular views …

AI: IN GAMES / REAL WORLD

4

In the real world …

2016 Libratus wins $1.7m in chips at Texas Hold’em AI system from CMU has beaten four of the world’s best poker players in a 20-day tournament. 2016-2017 AlphaGo beats world’s top Go players AI system from DeepMind /Google CMU has beaten Go’s grandmasters in 3-games matches

5

GAMES AND AI

AI: IN THE REAL WORLD

6

In the real world …

2014 Alexa: virtual personal assistant, smarthome 2016 J. Bezos’ WP news writer, since Rio Olympics

AI: IN THE REAL ROBOTIC WORLD

7

In the robot world …

AI: IN THE DIGITAL WORLD

8

In the daily digital life …

HOW DO WE LABEL X AS AN AI?

9

The science of making machines do things that would require intelligence if done by man (Bertram Raphael) Many different views, but all share a common concept … Intelligence ?

INTELLIGENCE, OPERATIONAL

10

(Intelligence) The cognitive ability of an individual (entity) to learn from experience, to reason well, to remember important information, and to (effectively) cope with the demands of daily living (Robert Sternberg) Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment (Nils J. Nilsson, The Quest for Artificial Intelligence, 2009)

ARTIFICIAL INTELLIGENCE

11

Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment (Nils J. Nilsson)

Is quicksort an example of AI? Is a pocket calculator an example of AI? q Scale q Autonomy q Speed q Flexibility q Generality q Capability q …. Intelligence ~ Human-related notion

THE BIRTH OF AI (1956)

§ Participants included Marvin Minsky, John McCarthy, Claude Shannon, Ray Solomonoff, Arthur Samuel, Allen Newell, Herbert Simon

§ “We propose that a two-month, ten-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, NH. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”

12

APPROACHES TO AI

13

The science of making machines that:

Think like people Act like people Think rationally Act rationally

Thought processes and reasoning Behavior Fidelity to human performance Ideal performance Functionally equivalent, imitation: it doesn’t matter how. Turing Test Theory of the mind ⟷ AI computer models, Cognitive science An agent that does the “right thing” based

• n what it knows (the thought process

doesn’t matter). Rational agents Laws of thought to represent and reason about things: Logics 973537498401 ?

STRONG VS. WEAK AI

– Weak AI: An artificial intelligence system can (only) act like it thinks and has a mind

14

Philosopher J. Searl (1980) made a distinction between different hypothesis about AI Strong AI: An artificial intelligence system can think and have a mind

R&N, Ch. 26

NARROW VS. GENERAL AI

– Narrow AI: An artificial intelligence system that replicates and maybe surpasses human intelligence for a dedicated purpose. “Applied” AI to a “narrow” domain.

15

General AI: A general-purpose artificial intelligence system that replicates and maybe surpasses human intelligence. A system with comprehensive knowledge and cognitive capabilities. Processing speed and capacity can be larger than humans. Embodiment (?) → Strong AI

GENERAL AI AND AI SAFETY

Elon Musk: AI is “our greatest existential threat.” Stephen Hawking: “Success in creating AI would be the biggest event in human

• history. Unfortunately, it might also be the

last...” Bill Gates: “First, the machines will do a lot

• f jobs for us and not be super intelligent.

That should be positive if we manage it

• well. A few decades after that, though,

the intelligence is strong enough to be a concern.”

16

THE TECHNOLOGICAL SINGULARITY

Emergence of superhuman intelligence Key idea: self-improvement Source of name: analogy between inability to predict events after the development of a superintelligence, and the space-time singularity beyond the event horizon of a black hole Some predict: this century Others argue: never

17

18

19

AI: DATA SCIENCE + SUPERCOMPUTING

20

§ Data, data, data… to learn from!

§ Sensors § Internet § Social nets § Clouds § Smartphones

§ Fast computers § GPUs § New techniques http://www.nature.com/nature/journal/v518/n7540/abs/nature14236.html

THIS COURSE: AI ~ ACT RATIONALLY

21

Think like people Act like people Think rationally Act rationally

Thought processes and reasoning Behavior Fidelity to human performance Ideal performance Functionally equivalent, imitation: it doesn’t matter how. Turing Test Theory of the mind ⟷ AI computer models, Cognitive science An agent that does the “right thing” based

• n what it knows (the thought process

doesn’t matter). Rational agents Laws of thought to represent and reason about things: Logics

RATIONAL DECISION-MAKING

22

This course: “Solving problems in the best possible way!” = Computational Rationality + (Modeling + Problem Solving)

= Mathematical and computational techniques for rational decision-making ↔ Act rationally

§ Provably optimal § Provably sub-optimal § Heuristic

Fundamental question: How to issue the sequence (≥ 1) of decisions that best/optimally (given accessible information and data) achieve my performance objectives?

DECISION-MAKING

23

… a single decision, or a sequence of decisions ...

RATIONALITY?

24

Not to be absolutely certain is, I think,

• ne of the essential things in rationality

(Bertrand Russel) Definitions are the guardians of rationality, the first line of defense against the chaos of mental disintegration (Ayn Rand) In everything, one thing is impossible: rationality (Friedrich Nietzsche)

AGENT RATIONALITY, OPERATIONALLY

25

1. Representation of the Problems / World Modeling!!! 2. Declare Goals and Preferences → Performance measure: Objective criterion to asses degree of success

• 3. Use all the necessary

available information (from sensors+built-in)

• I. Newton

4. Make decisions (act) that, given (1+2+3), maximize the expected performance (with formal guarantees)

More things should not be used than are necessary (Ockham)

Rationality ≠ Omniscience

LEARNING OBJECTIVES

26

1. How to make effective abstractions in order to categorize problems and define efficient problem representations 2. How to define goals, preferences, and utilities to effectively direct the problem solving process

Problem Formal Model Abstraction Problem Class My goal is to get to West-Bay (possibly in the minimum time) I don’t have a specific goal, but different states have a different utility (numeric value) I have different preferences

LEARNING OBJECTIVES

27

• 3. How to use all the available information to compute, or

learn, optimal decisions: defining how to act

Closed loop control State A State B Action a

Policy: (state → action) → maximize utility

Open loop control

Plan: (action a, action b, action c, … action n) → goal Classification: pattern → class Plan: (𝑦, = 𝑏, 𝑦/ = 𝑐, 𝑦1 = 𝑑, … 𝑦3 = 𝑜) → goal Regression: observed input → predicted output

Learning a discrete mapping Learning a function

(RATIONAL) AGENT

28

Agent Function, for a Rational Agent: For each possible input, the agent selects an action (1) that is expected to maximize the performance measure (2), given the evidence provided by the percept sequence (3) if need, and whatever built-in knowledge the agent has (4)

Agent Sensors Actuators Environment Percepts Actions

?

States of the representation: feasible configurations expressed in terms of selected variables of interest

§ Rational: maximally achieving pre-defined goals (based on representation, accessible data and information) § Rationality only concerns what decisions are made (not the thought process behind them)

EXAMPLE: VACUUM-CLEANER WORLD

29

§ Elements: Vacuum-cleaner robot, dirt, two (interconnected) rooms § Actions: Left, Right, Suck, NoOp § World States: 8 feasible configurations

VACUUM-AGENT FUNCTION

30

§ Actions: Left, Right, Suck, NoOp § Percepts: Location and Status, e.g., [A, Dirty] function Vacuum-Agent ([location, status]) returns an action

• if status = Dirty then return Suck
• else if location = A then return Right
• else if location = B then return Left

Is the agent behavior rational?

RATIONALITY VS. GOAL, INFORMATION

31

function Vacuum-Agent ([location, status]) returns an action

• if status = Dirty then return Suck
• else if location = A then return Right
• else if location = B then return Left

§ Environment: Static § Goal: Suck all the dirt in the minimum time § Environment: Static § Goal: Maximize collected reward {+1 dirt cleaning, -1 moving} § Environment: Dynamic § Goal: Maximize collected reward {+1 dirt cleaning, -1 moving} § Environment: Static, dirt is placed at the second time step § Location sensor: Unreliable, except for the first location § Goal: Suck all the dirt in the minimum time

TASK ENVIRONMENT ~ PROBLEM

32

Agent Sensors Actuators Environment Percepts Actions

? Agent behavior: Agent function that maps percept sequences to an action (we will learn how to design it for a rational agent!) Task environment (~Problem): Environment + Actuators + Sensors + Performance A vast range of task environments arise in AI …

EXAMPLE TASK ENVIRONMENTS

33

Chess with clock Taxi (self-)driving

PROPERTIES OF TASK ENVIRONMENTS

34

Ø Fully observable vs. Partially observable

• Full observability means access to the complete state: all

aspects that are relevant (necessary) to the choice of an action in relation to the selected performance measure Ø Deterministic vs. Stochastic

• If current state + executed action completely determine the next

state, then the environment is deterministic.

• Partial observability requires to treat the environment as

stochastic

• → Uncertainty derives from partially observable and/or not

deterministic environments Ø Known vs. Unknown

• In a known environment, all the outcomes (or outcome

probabilities) of an action are given (the agent knows the “rules”)

• In an unknown environment the agent has to learn how it works

PROPERTIES OF TASK ENVIRONMENTS

35

Ø Episodic vs. Sequential

• Agent’s experience is a collection of episodes (percept + action),

and episodes are independent from each other.

• In sequential environments, each decision could affect all future

decisions (the agent needs to think ahead) Ø Static vs. Dynamic

• The environment is dynamic if it changes over time
• It’s semi-dynamic if the agent’s performance score changes

Ø Discrete vs. Continuous

• It applies to environment states, to time, to percepts and actions

Ø Single agent vs. Multi-agent

• It can be a competitive or a cooperative environment based on

the impact they have on each other’s performance

EXAMPLES OF TASK ENVIRONMENTS

36

Different task environments require different techniques for problem representation, agent design, and rational decision-making

EXAMPLES OF TASK ENVIRONMENTS

37

Different task environments require different techniques for problem representation, agent design, and rational decision-making

BIG PICTURE

38

Problem Representation: states, transitions Constraint Satisfaction models Policy-based sequential decision-making in stochastic environments Sequential action learning through direct interaction with stochastic unknown environments Games: Decision-making in multi-agent adversarial environments Deep learning architectures Continuous and discrete

• ptimization

for decision-making Automated planning in fully observable and deterministic environments Supervised learning of regression models and classifiers Modeling and prediction of Markov random processes Probabilistic inference

39

40

LEARNING OBJECTIVES

41

By the end of the course, you should have learned a number of different techniques for decision-making and learning, and you should be able to: 1. Identify the type of an AI problem. 2. Formulate the problem as a particular type. 3. Identify the most appropriate technique to use for decision- making. 4. Compare the difficulty of different versions of AI problems, in terms of computational complexity and the efficiency of existing algorithms. 5. Implement, evaluate and compare the performance of various AI algorithms. Evaluation could include empirical demonstrations and/or theoretical proofs.

COURSE LAYOUT AND INFO

42

§ Based on lectures § Reference to textbooks and additional material will be provided topic by topic, however, everything you need will/should be covered by lecture handouts § 10 homework assignments: theory and programming (in Python + use of the Anki Cozmo robot) § One midterm + One final § Grading: 35% Final, 15% Midterm, 50% Homework § No smartphones or laptops (when not used for taking notes)

QUESTIONS?

43

Piazza: used for questions, announcements, and polls My office: office hours + drop in at any time! (or send an email)

Course information (contents, schedule, policies…) :

https://www.qatar.cmu.edu/~gdicaro/15381/